sintel-dev/Orion

Unsupervised time series anomaly detection library

What it solves

Orion is designed to identify rare patterns and anomalies in time series data without requiring labeled training sets. It provides a collection of verified machine learning pipelines that automate the process of flagging unusual signals for expert review.

How it works

The library implements various unsupervised learning architectures to detect anomalies. It offers multiple "verified" pipelines, including those based on Generative Adversarial Networks (GANs), Autoencoders (Dense, VAE, and LSTM), and other specialized methods like AER (Auto-Encoder with Regression) and Matrix Profile. Users can fit a pipeline to training data and then use the detect method to identify anomalies in new incoming signals, returning the start, end, and severity of each detected event.

Who it’s for

This tool is for data scientists and researchers who need to monitor time series signals and detect outliers or rare events automatically using a variety of state-of-the-art unsupervised ML models.

Highlights

  • Multiple Pipeline Options: Includes a wide array of models such as TadGAN, AER, and AnomalyTransformer.
  • Automated ML Tools: Built using automated machine learning tools from the MIT Data to AI Lab.
  • Benchmarked Performance: Maintains a leaderboard comparing pipeline performance against a baseline (ARIMA) across 12 different datasets.
  • Simple API: Provides a straightforward fit and detect workflow for rapid deployment.